|
| Reference | Analysis type | Feature extraction | Classification | Highlights |
|
| [145] | Vibration | Ensemble empirical mode decomposition (EEMD) and correlation coefficient (CC) along with singular value decomposition (SVD) | (1) The single probabilistic classifiers
(2) The single probabilistic and Bayesian machines
(3) Pairwise-coupled
(4) Two classifiers without pairwise-coupling | (i) It diagnoses multiple and single faults
(ii) There is simultaneous fault diagnosis
(iii) The accuracy for a single fault is 92.62% and for simultaneous faults is 85.73% |
| [146] | Vibration and acoustic emission | Unsupervised feature | Dictionary learning | (i) There is online monitoring
(ii) There are different operational conditions
(iii) There are good computational costs |
| [147] | Vibration | Multiscale entropy | SVM | (i) It diagnoses multiple faults
(ii) The average accuracy is 96.25% |
| [148] | Vibration | Hilbert transform (HT) and continuous wavelet transform (CWT) | Neural network (multilayer perceptron), neural network (radial basis function), and support vector machines | (i) There is multiclass FDD
(ii) SVM is found to be the best (with SVM 99.71%) compared to NN classifiers |
| [149] | Vibration and current | SVM bootstrap based technique with features data fusion | Kernel-nonlinear SVM along with Gaussian radial basis function | (i) SVM multiclassification scheme is presented
(ii) It diagnoses multiple faults
(iii) There are different operational conditions
(iv) The average accuracy is 99.4% |
| [150] | Vibration and current | Statistical features analysis | SVM and multiclass SVM | (i) It diagnoses multiple faults
(ii) There is electrical and mechanical faults’ prediction
(iii) There are different operational conditions
(iv) MSVM showed an ability of predicting all mechanical faults
(v) SVM is better than MSVM for electrical faults diagnosis
(vi) The average accuracy is 93.28% |
| [154] | Vibration | Deep learning | Deep belief networks (DBN) | (i) There is automatic FDD
(ii) The proposed approach could detect the fault directly from frequency distribution without needing traditional feature extraction methods
(iii) It learns multiple layers of representation and models high-dimensional data
(iii) The average accuracy is 99.00% |
| [155] | Vibration | Deep learning | Sparse autoencoder | (i) It diagnoses multiple faults
(ii) It prevents training process overfitting
(iii) The average accuracy is 97.61% |
| [156] | Infrared thermal (IRT) images | Deep learning | Convolutional neural networks | (i) There is online monitoring
(ii) There are different operational conditions
(iii) The average accuracy is 95% |
| [157] | Stator current | Deep learning | Deep neural network | (i) IM bearings monitoring tool based on deep learning is proposed
(ii) Different load conditions 25%, 50%, 75%, and 100% are tested
(iii) Deep neural network showed better classification accuracy than shallow neural network (SNN) and principle component analysis (PCA) |
| [158] | Vibration | Kurtogram and deep learning | Recurrent NN, long-/short-term memory, and gated recurrent unit | (i) FDD method based on kurtogram and deep learning is proposed
(ii) Computational time, computing resources and number of layers, is small
(iii) Misclassification occurred
(iv) The average accuracy is 98% |
| [159] | Vibration | Neural networks | Transfer learning | (i) Bearing FDD approach based on transfer learning with neural networks is proposed
(ii) Different working conditions are analysed
(iii) Training time comparing with NN is reduced
(iii) It deals with massive data
(iv) Transfer learning improved the classification accuracies
(v) The total classification accuracy is improved by 10.4 % |
| [160] | Acoustic emission | Transfer learning-based convolutional neural network | Transfer learning | (i) Bearing FDD acoustic spectral imaging and transfer learning under variable speed conditions and different rotational speeds is proposed
(ii) Two-dimensional acoustic frequency spectral imaging with a transfer learning is discussed
(iii) The proposed method achieved an average accuracy of 94.67% |
| [161] | Vibration | Long-/short-term memory recurrent neural network and feature-transfer learning (joint distribution adaptation) | Grey wolf optimization algorithm | (i) Bearing FDD based on adaptive deep transfer learning is proposed
(ii) Massive labeled fault data is collected and analysed
(iii) The proposed method achieved an average accuracy of 99.4 % |
| [162] | Vibration | Multiscale deep intraclass adaptation network | Multiple scale feature learner | (i) Bearing FDD is based on multiscale deep intraclass transfer learning
(ii) Different working conditions are analysed
(iii) The proposed method achieved an average accuracy of 99 % |
| [163] | Vibration | Hybrid deep signal processing approach | Autoencoder | (i) Deep learning with time synchronous resampling mechanism is proposed
(ii) The proposed method dealt with shift variant properties, periodic inputs, and misclassification challenges
(iii) The proposed method achieved an average accuracy of 99 % |
|
